CapitalGuard Observatory

CHECK WHAT YOUR AI CAN REACH.

Choose your tool. Get a clear risk snapshot, then run a quick check.

Choose my tool

Quick check

Which AI tool do you use?

ChatGPT

Your quick snapshot

Review access

May reach

prompts and uploaded files

Check first

Data Controls and model-improvement choice

Run my quick check View evidence

Reproducible benchmark

Reproduce detection and runtime enforcement, case by case.

The repository benchmark measures six exposure classes. Enforced Mode follows 23 runtime cases. The delegation benchmark tests twelve signed handoff boundaries. The intent benchmark tests fifteen signed-plan, action-target, sequence, replay, and uncertain-outcome cases.

Runtime benchmark OWASP evidence map Scanner benchmark Actions benchmark License benchmark Permission drift benchmark MCP integrity benchmark Delegation boundary benchmark Intent conformance benchmark Managed-session stop-path evidence Evidence index